Transformer-based magnetic resonance-to-computed tomography synthesis for radiotherapy for cervical cancer: model development and clinical evaluation

Integrating magnetic resonance (MR) imaging (MRI) with computed tomography (CT) for radiotherapy in cervical cancer management enhances soft tissue visualization; however, it poses challenges such as registration errors and high costs. Therefore, this study introduces CervixAttenNet, a transformer-based MR-to-CT synthesis model that leverages queryselected attention to improve contrastive learning and achieve precise MR-to-synthetic CT (sCT) image alignment. We analyzed the data of 100 patients with cervical cancer who underwent intensity-modulated radiotherapy. Preprocessed MRI and planning CT (pCT) images were registered, with CT serving as the ground truth for sCT image generation. The architecture of CervixAttenNet optimizes feature alignment across domains using global attention mechanisms. Image quality and dosimetric accuracy were rigorously evaluated. CervixAttenNet achieved a MAE of 118.82 HU, RMSE of 84.51 HU, PSNR of 23.14 dB, and SSIM of 0.841, demonstrating significantly improved synthesis performance over CycleGAN. Dosimetric differences between sCT and pCT were minimal, and similarly small mean dose differences for organs at risk. Gamma analysis further confirmed the high agreement between sCT-and pCT-based dose distributions, with consistently high passing rates. CervixAttenNet is a promising tool for MR-guided radiotherapy planning in cervical cancer management, providing improved accuracy, reduced radiation exposure, and minimized registration errors.